💻 computer science

EpiLENS: Patient-Relative Epileptogenic Zone Localization from Multi-Center Intracranial EEG

EpiLENS is a novel patient-relative framework that improves the localization of epileptogenic zones in drug-resistant epilepsy by employing a Conservative Dual-Evidence Localization strategy to combine a patient-specific baseline deviation network with a boundary-coverage ranking network, demonstrating robust generalization across diverse clinical centers and recording conditions.

Original authors: Yuanchu Gong, Zibo Yan, Yibo Lyu, Chen Chen, Sixian Chan, Yalin Wang

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yuanchu Gong, Zibo Yan, Yibo Lyu, Chen Chen, Sixian Chan, Yalin Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain as a bustling, chaotic city. Usually, the traffic flows smoothly, but sometimes, a specific neighborhood gets stuck in a massive, uncontrollable gridlock. In the medical world, this is an epileptic seizure. For some people, medication can't clear the traffic; the gridlock keeps happening no matter what. The only way to fix it is for a surgeon to go in and remove that specific, broken neighborhood. But here's the tricky part: the "broken" neighborhood isn't always obvious. It's hidden deep inside the brain, and every person's city layout is different. To find it, doctors use tiny microphones (electrodes) placed directly on the brain to listen to the electrical chatter. The big challenge is figuring out which microphone is listening to the trouble and which ones are just hearing the normal noise of the city. If they remove the wrong neighborhood, the seizures keep coming. If they remove too much, they might hurt the patient's memory or movement. So, the goal is to be a super-precise detective, listening to thousands of tiny signals to pinpoint exactly where the trouble starts, even when the signals look different for every single patient.

This is where a new tool called EpiLENS comes in. Think of EpiLENS as a smart, patient-specific detective squad that stops trying to use a "one-size-fits-all" rulebook. Instead of asking, "Is this signal louder than the average person's signal?" (which is like judging a whisper in a library by how loud it is compared to a stadium roar), EpiLENS asks, "Is this signal acting weird compared to this specific person's own normal?"

The researchers built EpiLENS to solve a messy problem: data from different hospitals, different patients, and even different seizures in the same patient can look totally different. A signal that looks "abnormal" in one person might be "normal" in another. To handle this, EpiLENS uses a clever two-part strategy called Conservative Dual-Evidence Localization (CDEL). Imagine two detectives working on the same case but with different styles.

The first detective, PRQ-Net, is the "Steady Evaluator." It listens to the patient's brain over many different seizures and looks for patterns. It's great at spotting the usual suspects but has a safety net: it pays extra attention to signals that are weird in just some seizures, making sure it doesn't ignore a troublemaker just because they were quiet in a few instances. It establishes a baseline of what is "normal" for that specific person and flags anything that deviates from it.

The second detective, BCR-Net, is the "Boundary Hunter." It knows that the hardest part of the job is finding the fuzzy line between the "bad" neighborhood and the "good" one. It focuses on the channels that are right on the edge, trying to rank them so that the most suspicious ones rise to the top of the list. It's like a referee who is really good at spotting the players who are almost breaking the rules, ensuring the whole "bad" group gets identified, not just the obvious ones.

Here is the magic trick: EpiLENS doesn't let these two detectives argue or mix their rules while they are training. They train separately. But when it's time to make a final decision, they combine their notes using a strict, pre-set formula. The "Steady Evaluator" gets 80% of the vote, and the "Boundary Hunter" gets 20%. This ensures the final decision is mostly based on solid, repeated evidence, but still gets a helpful nudge from the expert who knows how to find the tricky edges.

The team tested this system on a huge group of patients from four different medical centers, involving 80 people, 256 seizures, and over 7,600 brain channels. They found that EpiLENS was better at finding the epileptic zone than older methods, including complex neural networks that tried to learn directly from raw brain waves. Crucially, it worked well even when the data came from a hospital the system had never seen before, proving that focusing on "what is normal for this patient" is a much stronger strategy than trying to find a universal rule for everyone. The results suggest that by treating each patient's brain as its own unique city and comparing signals only against that city's own history, doctors can get a clearer, more reliable map of where to operate, potentially leading to better outcomes for those with drug-resistant epilepsy.

Technical Summary: EpiLENS

Problem Statement

Drug-resistant epilepsy (DRE) requires precise localization of the epileptogenic zone (EZ) for successful neurosurgery. Current intracranial EEG (iEEG) methods face significant challenges in heterogeneous multi-center settings due to substantial variability across patients, seizure types, implantation layouts, recording systems, and clinical centers.

Existing approaches typically rely on globally trained, channel-wise classifiers. These methods often fail because they assume that comparable electrophysiological values retain comparable meaning across different patients. However, multi-center data reveals that absolute scores are brittle due to:

Patient Heterogeneity: Differences in recorded feature scales and the specific anatomical sites compared against each channel.
Seizure Heterogeneity: EZ evidence may recur consistently or appear strongly only in a subset of a patient's seizures.
Set Imbalance: Annotated EZ channels are typically a minority within the implanted set, causing average classification to be insensitive to whether the complete patient-specific target rises to the top.

Consequently, a global boundary often learns acquisition or implantation differences rather than patient-specific epileptogenic abnormalities. The core problem is determining whether a contact is abnormal relative to the other channels observed for the same patient, rather than against a universal cohort-wide value.

Methodology: EpiLENS

The authors propose EpiLENS, a patient-relative, primary-guided collaborative framework designed for cross-patient epileptogenic localization in heterogeneous multi-center iEEG. The framework operates on a Conservative Dual-Evidence Localization (CDEL) strategy, which asymmetrically combines two independently trained branches at inference.

1. Patient-Relative Evidence Construction

Before modeling, the system constructs patient-relative evidence:

Feature Extraction: Nine spectral and waveform descriptors are extracted per channel and temporal window.
Self-Referenced Views: For each channel, features are transformed into four views relative to the channel's own pre-onset baseline (derived from negative-time windows of the same seizure):
1. Absolute (ABS): Raw electrophysiology.
2. Difference ( $\Delta$ REF): Additive deviation from the baseline.
3. Standardized ( $Z$ - $\Delta$ ): Standardized deviation accounting for natural variability.
4. Log-Relative (LOG-R): Multiplicative change.
Normalization: These views are standardized within the patient's implanted set, ensuring no contact is judged against a universal scale.

2. The Dual-Branch Architecture

The framework employs two branches with identical encoder topology but independent parameters:

PRQ-Net (Patient-Relative Quantile Network):
- Role: Serves as the primary probabilistic classifier.
- Mechanism: Aggregates recurrent cross-seizure evidence. Crucially, it employs a bounded lower-tail correction (Q10). Instead of simple mean aggregation, which can suppress EZ-like activity appearing strongly in only a subset of seizures, PRQ-Net retains the 10th percentile of seizure-level scores.
- Output: A patient-relative probability of belonging to the Non-Epileptogenic Zone (NEZ).
BCR-Net (Boundary-Coverage Ranking Network):
- Role: Provides complementary ranking evidence.
- Mechanism: Focuses supervision on two specific failures: confusion near the EZ/NEZ boundary and incomplete recovery of the annotated EZ set.
  - Boundary Mining: Identifies the weakest annotated EZ channels and the most EZ-like NEZ channels to enforce separation.
  - Coverage: Penalizes annotated EZ channels that fail to enter the top-ranked positions for that patient.
- Output: An EZ-oriented logit used for ranking.

3. Conservative Dual-Evidence Localization (CDEL)

At inference, the branches are fused using a fixed, asymmetric rule:
$p^{CDEL}_{ic} = 0.8 \cdot p^{PRQ}_{ic} + 0.2 \cdot p^{BCR}_{ic}$
Where $p^{PRQ}$ is the primary NEZ probability and $p^{BCR}$ is the complementary probability derived from the BCR-Net logit. This rule ensures PRQ-Net remains the primary driver while BCR-Net contributes patient-specific ranking evidence without overriding the primary classifier. The branches are trained independently with no joint loss function.

Key Contributions

EpiLENS Framework: A patient-relative, primary-guided collaborative framework for cross-patient EZ localization in heterogeneous multi-center iEEG, addressing the limitations of global classifiers.
CDEL Strategy: A novel fusion mechanism that combines PRQ-Net's stable patient-relative quantile evidence with BCR-Net's boundary- and coverage-aware ranking through a fixed, conservative fusion rule.
Comprehensive Evaluation: Rigorous validation on a heterogeneous four-center cohort (80 patients, 256 seizures, 7,635 channels) using patient-disjoint cross-validation, cross-seizure tests, and leave-one-center-out generalization.

Experimental Results

The framework was evaluated against classical feature-based methods (Logistic Regression, SVM, Random Forest), raw-iEEG neural baselines (SEEGNet, TimeConv-CNN, CLAP, SEEGformer), and patient-relative controls.

Performance: EpiLENS (CDEL) achieved the highest Macro-F1 (0.6371) and EZ-F1 (0.4485), outperforming the strongest baseline (Logistic Regression with patient-wise z-score) by +0.0306 in Macro-F1 and +0.0497 in EZ-F1.
Bias Reduction: The method achieved near-zero EZ-Fraction Bias, indicating that improvements were not due to systematically enlarging or shrinking the predicted EZ set.
Component Analysis:
- Patient-relative normalization was identified as the dominant component; removing it reduced Macro-F1 from 0.6282 to 0.5550.
- The BCR-Net's boundary and coverage objectives provided complementary ranking gains, improving EZ-AUPRC and NDCG-EZ significantly over PRQ-Net alone.
Generalization:
- Cross-Seizure: Performance improved with the inclusion of more seizures (1 vs. 2 vs. all), confirming the value of aggregating repeated recordings.
- Leave-One-Center-Out (LOCO): CDEL achieved the highest center-mean Macro-F1 (0.6163) and the strongest worst-center result (0.5773) across four centers, demonstrating robust transferability to unseen clinical centers.

Significance and Claims

The paper claims that patient-relative modeling is an effective strategy for cross-patient EZ localization. By shifting the reference frame from a universal cohort-wide value to a patient-specific internal baseline, EpiLENS aligns spatial localization with the way implanted evidence is interpreted in clinical practice.

The authors emphasize that the proposed design consistently improves localization while remaining robust to variations in:

Seizure availability (cross-seizure aggregation).
Recording centers (LOCO generalization).
Implantation and device heterogeneity.

The study concludes that the combination of stable patient-relative probabilities (PRQ-Net) and complementary boundary/coverage ranking (BCR-Net) offers a coherent mechanism for handling the inherent variability in multi-center iEEG data, without requiring complex connectivity features or raw-signal representations that proved sensitive to recording differences in this context.